AgentSkillsCN

memory-workflow

git-review的数据工作流、内存系统分离，以及观察规则。明确指定在何种场景下使用哪种内存系统，并避免重复存储。适用于保存决策、记录观察，或管理跨会话状态的场景。

SKILL.md
--- frontmatter
name: memory-workflow
description: "Data workflow, memory system separation, and observation rules for git-review. Defines which memory system to use for what, and how to avoid duplication. Use when saving decisions, recording observations, or managing cross-session state."

Memory & Data Workflow

Data Workflow (ENFORCED)

  • •Linear = single source of truth for requirements, status, and acceptance criteria
  • •claude-mem = cross-session decision memory (supplements Linear, never duplicates)
  • •Local plan files = ONLY the planner agent writes these (plans/<feature>-plan.md), for code-level implementation detail too granular for Linear. All other agents write to Linear comments.
  • •Requirements, specs, critiques, and review results go to Linear comments, not local files

Memory System Separation

SystemPurposeWhat to Store
claude-memHuman-drivenUser preferences, architectural decisions, debugging insights, cross-session learnings
Claude-FlowAgent-drivenAgent coordination patterns, code optimization history, model performance data
SerenaCode-drivenProject code structure, conventions, build commands (auto-discovered via onboarding)
LinearProject managementTicket requirements, acceptance criteria, status tracking

No overlap: Each memory system serves a distinct purpose; never duplicate content across systems.

claude-mem 3-Layer Workflow

  1. •search(query) — Get compact index with IDs (~50-100 tokens/result)
  2. •timeline(anchor=ID) — Get chronological context around interesting results
  3. •get_observations([IDs]) — Fetch full details ONLY for filtered IDs

NEVER skip to get_observations without filtering first (saves ~10x tokens).

When to Save (claude-mem)

  • •User states a preference or working style
  • •An architectural or design decision is made
  • •You discover a non-obvious project convention or pattern
  • •A debugging session yields a reusable insight
  • •A mistake is made that others (or future you) should avoid

When NOT to Save

  • •Transient debugging context (one-off stack traces, temp workarounds)
  • •Information already in CLAUDE.md, README, or docs
  • •Obvious facts derivable from reading the code

Observation Rules

  • •Memory observer agents are DEPRECATED — use /checkpoint skill instead
  • •If observers must be used: only record what was actually observed. Never fabricate details.
  • •If a session has minimal activity, record that limitation explicitly rather than padding with assumptions

MCP Server Usage

  • •claude-mem: Cross-session memory only. Save decisions/patterns (brief). NEVER duplicate Linear content.
  • •Linear: Source of truth for requirements, status, and acceptance criteria. ALL tickets tracked there.